Security of Edge AI Against Hardware Attacks

Explore the security challenges of Edge AI against hardware attacks in this tinyML Talks Germany meetup webinar. Delve into the vulnerabilities of neural networks, particularly in edge AI applications where physical access to devices poses additional risks. Learn about potential attack vectors like side-channel analysis and fault attacks, and understand how attackers might attempt to reverse engineer and copy neural networks. Gain insights into countermeasures, including masking techniques, to protect valuable intellectual property. Discover the intricacies of network structure, neuron parameters, power traces, and activation functions in the context of security. Examine the effectiveness of differential power analysis and the implications of increasing traces on retrieving weights. Consider the scalability of attacks with network size and the impact of prior knowledge on attack success. Engage with discussions on hardware counter measures, parallel implementation strategies, and the generation of adversarial examples. Conclude with a comprehensive overview of network security in Edge AI and participate in an interactive Q&A session addressing audience queries.

Introduction
Overview
Side channel analysis
Differential power analysis
Fault attacks
Neural Network
Network Structure
Neuron Parameters
Power Trace
Activation Function
Retrieving Weights
Increasing Traces
Results
Counter measures
Masking
Takeaways
Questions
Thank you
Poll
Q1 How many neurons do the mentioned MLCN networks contain
How many neurons do the mentioned MLCN networks contain
How well does it scale with the network size
Does it make any difference
Generating adversarial examples
IP theft
Least negative impact
Hardware counter measures
How successful is an attack
Prior Knowledge
Random Input
Retrieve Network
Network Security
Parallel Implementation
Noise
Other attacks
Summary
Audience questions
Sponsors